Research Question

How do explainable AI methods (SHAP or saliency maps) applied to radiomics models for brain-tumour radiotherapy response affect clinician trust and adoption, and do they reveal biologically meaningful features linked to hypoxia or necrosis?

Jana M
Created at November 17, 2026

AI Novelty Assessment

7/10

High Novelty

This research question explores a largely uncharted area with significant potential for new discoveries.

Detailed Analysis

The question combines three elements that do not appear to have been studied together. Searches across Semantic Scholar, PubMed, OpenAlex, Europe PMC and arXiv returned general AI and radiomics reviews for neuro-oncology and glioma, and a few radiogenomic studies linking MRI radiomic features to molecular programs (for example, IvyGAP-based subcompartment analysis in 2026 and spherical radiomics in 2026). These show that radiomic features can carry biological meaning, which is the premise for the question's biological-plausibility test. However, I found no study that applies SHAP or saliency methods to radiomics models for brain-tumour radiotherapy response, measures clinician trust or adoption, or links the explanations specifically to hypoxia or necrosis. Explainable AI is discussed in reviews of other cancers, which suggests the methodological interest exists but has not reached this clinical setting. The PubMed search returned no results and arXiv returned none directly relevant, so coverage is thinner than the other databases. The gap is substantial, which is why the score is high.

Related Academic Papers

6 papers found relevant to this research question. Each paper is scored by how closely it relates to the question.

6/10Relevance
0 citations

Abstract

Background: Glioblastoma exhibits profound intratumoral heterogeneity, with anatomically distinct tumor zones characterized by divergent molecular programs that drive therapy resistance. Whether magnetic resonance imaging (MRI)-derived radiomic features can capture these regional transcriptomic differences remains unknown. We aimed to determine whether subcompartment-level radiomic features associate with transcriptomic pathway enrichment scores derived from biologically approximate tumor zones. Methods: We matched 28 patients (mean age 58.5 years; 13/28 MGMT methylated) across the IvyGAP RNA-seq atlas and the IVYGAP-RADIOMICS datasets. Single-sample GSEA (ssGSEA) pathway scores were computed for 24 gene sets. Radiomic features (3920 per subcompartment) were reduced to 597. Nested leave-one-patient-out cross-validation (LOPO-CV) with Elastic Net served as the primary predictive analysis; linear mixed-effects models (LMM) provided exploratory associational analysis. Analyses used a biologically motivated but spatially non-co-registered zone-to-subcompartment mapping; all reported associations are zone-approximate. Results: Twenty-one of 24 pathways showed no predictive signal (R2cv ≤ 0). Inflammatory Response (R2cv = 0.185, 95% CI [0.071, 0.355], p = 0.008) was the only pathway supported by both the nested CV (FDR = 0.096) and the exploratory LMM (FDR = 0.024, ΔR2 = 0.214 beyond subcompartment effects) analyses; the LMM association was robust to clinical covariate adjustment (likelihood ratio test p = 0.004). Angiogenesis (R2cv = 0.209, 95% CI [0.028, 0.353], p = 0.006) reached nested CV significance (FDR = 0.096) but was not corroborated by the LMM (FDR = 0.445); it is therefore reported as a tentative single-framework signal requiring independent validation. T2-derived texture features were selected in 100% of folds for both pathways. Conclusions: Inflammatory Response is the only pathway supported by both analytical frameworks; Angiogenesis is a tentative nested-CV-only signal pending independent validation. The absence of signal for 21 of 24 pathways should not be interpreted as evidence of biological inaccessibility: at N = 28 (vs. N ≈ 240 required by Riley criteria), severe underpowering, attenuation from the non-spatial zone-to-subcompartment mapping, and methodological constraints each independently suffice to suppress real associations. Five of the 24 gene sets (the IvyGAP zone modules) are non-independent from the outcome data and cannot be interpreted as discovery. All reported associations are zone-approximate and may partly reflect macro-compartment (between-subcompartment) effects; validation in larger cohorts with spatially precise co-registration is essential.

Why this paper is relevant

2026 study linking MRI radiomic features of tumour subcompartments to angiogenic and inflammatory transcriptomic programs in glioblastoma (IvyGAP). Supports the idea that radiomic features can carry biology, a precondition for the question's biological-plausibility test, but it does not use explainability methods or radiotherapy response.

Artificial intelligence for brain cancer management: multimodal digital workflows for diagnosis, treatment planning, and longitudinal monitoring.

G. Saibaba, Olawumi Giwa, T. Booth, Terézia Kisková-Šimková, Arban Uka, Florenc Skuka, B. Javed, Dário Ferreira, Sandra S. Ferreira, Jewel Sengupta, Muhammet Karaman, G. Calvo, Xin-Zhong Li, Tatjana Mitrović, A. Occhipinti, C. Angione, N. Missaoui, Milana Frenkel-Morgenstern (2026)

5/10Relevance
0 citations

Abstract

Brain cancers, especially glioblastoma, remain among the deadliest adult cancers, with outcomes largely unchanged despite multimodal treatments. This review summarizes cutting-edge artificial intelligence (AI) and machine learning (ML) advances transforming neuro-oncology in diagnostics, molecular profiling, treatment planning, and monitoring. Key findings show AI-driven radiomics and deep learning (DL) reaching over 90% accuracy in tumour segmentation and grading from MRI and whole-slide images, non-invasive IDH/MGMT prediction through liquid biopsy (LB) analysis, and augmented reality-guided resection that maximizes tumour removal while safeguarding expressive cortex. Treatment planning benefits from hybrid Convolutional Neural Networks (CNN)-Transformer models for immunotherapy stratification and blood-brain barrier penetrant drug repurposing, while real-time progression detection via multimodal integration helps differentiate true progression from pseudoprogression. Despite these advances, significant challenges remain, including data scarcity and imbalance in rare subtypes, domain shift due to imaging variability, black-box model behaviour eroding trust, regulatory requirements for prospective validation, and workflow fragmentation. Emerging solutions include federated and transfer learning for scalable model development, explainable AI (such as SHapley Additive exPlanations (SHAP) and attention interpretation for vision transformers) to foster clinician-AI collaboration, and foundation models pretrained on large-scale neuro-oncology datasets to facilitate personalization. Achieving this potential will depend on harmonized multi-omics registries, strong ethical and regulatory governance, and deliberate human-AI collaboration frameworks to integrate these tools into precision neuro-oncology.

Why this paper is relevant

2026 review of AI and ML across neuro-oncology diagnostics, molecular profiling, treatment planning and monitoring. Shows the field is active and may mention explainability in passing, but it does not study SHAP or saliency methods, clinician trust, or radiotherapy response specifically.

Radiomics in glioma: emerging trends and challenges

Zihan Wang, Lei Wang, Yinyan Wang (2025)

5/10Relevance
41 citations

Abstract

Radiomics is a promising neuroimaging technique for extracting and analyzing quantitative glioma features. This review discusses the application, emerging trends, and challenges associated with using radiomics in glioma. Integrating deep learning algorithms enhances various radiomics components, including image normalization, region of interest segmentation, feature extraction, feature selection, and model construction and can potentially improve model accuracy and performance. Moreover, investigating specific tumor habitats of glioblastomas aids in a better understanding of glioblastoma aggressiveness and the development of effective treatment strategies. Additionally, advanced imaging techniques, such as diffusion-weighted imaging, perfusion-weighted imaging, magnetic resonance spectroscopy, magnetic resonance fingerprinting, functional MRI, and positron emission tomography, can provide supplementary information for tumor characterization and classification. Furthermore, radiomics analysis helps understand the glioma immune microenvironment by predicting immune-related biomarkers and characterizing immune responses within tumors. Integrating multi-omics data, such as genomics, transcriptomics, proteomics, and pathomics, with radiomics, aids the understanding of the biological significance of the underlying radiomics features and improves the prediction of genetic mutations, prognosis, and treatment response in patients with glioma. Addressing challenges, such as model reproducibility, model generalizability, model interpretability, and multi-omics data integration, is crucial for the clinical translation of radiomics in glioma.

Why this paper is relevant

2025 review of radiomics in glioma covering emerging trends and challenges, including deep-learning integration. Identifies interpretability-related challenges in general terms, but does not evaluate explanation methods or adoption outcomes.

Artificial intelligence-based MRI radiomics and radiogenomics in glioma

Haiqing Fan, Yilin Luo, Fang Gu, Bin Tian, Yongqin Xiong, Guipeng Wu, Xin Nie, Jing Yu, Juan Tong, Xin Liao (2024)

5/10Relevance
132 citations

Abstract

The specific genetic subtypes that gliomas exhibit result in variable clinical courses and the need to involve multidisciplinary teams of neurologists, epileptologists, neurooncologists and neurosurgeons. Currently, the diagnosis of gliomas pivots mainly around the preliminary radiological findings and the subsequent definitive surgical diagnosis (via surgical sampling). Radiomics and radiogenomics present a potential to precisely diagnose and predict survival and treatment responses, via morphological, textural, and functional features derived from MRI data, as well as genomic data. In spite of their advantages, it is still lacking standardized processes of feature extraction and analysis methodology among different research groups, which have made external validations infeasible. Radiomics and radiogenomics can be used to better understand the genomic basis of gliomas, such as tumor spatial heterogeneity, treatment response, molecular classifications and tumor microenvironment immune infiltration. These novel techniques have also been used to predict histological features, grade or even overall survival in gliomas. In this review, workflows of radiomics and radiogenomics are elucidated, with recent research on machine learning or artificial intelligence in glioma.

Why this paper is relevant

Review of AI-based MRI radiomics and radiogenomics in glioma. Covers the radiomics-genomics biology link that explainability could expose, but focuses on diagnosis and genotype prediction rather than explainable radiotherapy response models or clinician trust.

4/10Relevance
2 citations

Abstract

Background: To develop and validate a novel spherical radiomics framework for predicting key molecular biomarkers-including MGMT promoter methylation, EGFR, and PTEN mutation status-and survival in glioblastoma (GBM) patients using multiparametric MRI. Methods: Using the UCSF-PDGM cohort, we propose a spherical radiomics framework in which tumor-centered concentric shells are generated at increasing radial distances from the tumor centroid and mapped onto two-dimensional surfaces for feature extraction. Radiomic features-including shape, first-order statistics, and texture descriptors (GLCM, GLRLM, GLDM, GLSZM, and NGTDM)-were extracted from 299 GBM patients using PyRadiomics across 4 tumor regions: necrotic core, T1-weighted contrast-enhancing region, T2/FLAIR hyperintense lesion, and a 2 cm peritumoral expansion region. Classification was performed using multiple machine learning models, including neural networks, logistic regression, random forest, and the Tree-based Pipeline Optimization Tool (TPOT). The proposed framework was further validated on an external UPENN-GBM cohort. Model interpretability was assessed using SHAP analysis, feature significance profiling, clustering visualization, and evaluation of radiomic patterns in relation to underlying biological processes. Radial transition analysis was conducted to quantify feature changes across adjacent tumor regions. Results: Spherical radiomics achieved AUCs of 0.82 for MGMT, 0.77 for EGFR, 0.76 for PTEN, and 0.79 for survival prediction, consistently outperforming conventional Euclidean radiomics (0.71, 0.61, 0.70, and 0.61, respectively). Consistent improvements were also observed in external validation based on UPENN-GBM cohort, with an AUC improved by 8% compared to Euclidean radiomics. GLCM-derived features were identified as the most informative predictors. Radial transition analysis using the Mann-Whitney U-test demonstrated that transition slopes between the T1-weighted contrast-enhancing and T2/FLAIR hyperintense regions, as well as between the T2/FLAIR hyperintense and peritumoral regions, were significantly associated with biomarker status. Conclusion: Radiomic features extracted from spherical surfaces at varying radial distances from the tumor centroid demonstrate stronger associations with key molecular markers and patient survival compared to conventional Euclidean radiomics, highlighting the value of spatially structured radiomic analysis for improved GBM characterization.

Why this paper is relevant

2026 spherical radiomics framework for glioblastoma radiogenomic biomarkers and survival. Shows radiomic features linked to molecular biomarkers, which bears on biological interpretation, but without explainability analysis or response outcomes.

4/10Relevance
0 citations

Why this paper is relevant

2026 review of AI radiomics and radiogenomics for pseudoprogression after glioblastoma immunotherapy and radiotherapy. Relevant to the response endpoint, but does not discuss SHAP, saliency, or clinician trust.

Generate your own research questions

ChatAcademia helps researchers discover novel research questions with AI-powered analysis.

Start Free Trial